outputs/accounting.json

{
  "table1": {
    "SGD": {
      "paper_total": 12,
      "pred_total_table1": 12,
      "pred_total_with_scales": 12,
      "match": true
    },
    "FlashSGD": {
      "paper_total": 6,
      "pred_total_table1": 6,
      "pred_total_with_scales": 6.0625,
      "match": true,
      "paper_total_grad_release": 4,
      "pred_total_grad_release_with_scales": 4.0625,
      "match_grad_release": true
    },
    "Adam": {
      "paper_total": 16,
      "pred_total_table1": 16,
      "pred_total_with_scales": 16,
      "match": true
    },
    "FlashAdam": {
      "paper_total": 7,
      "pred_total_table1": 7,
      "pred_total_with_scales": 7.125,
      "match": true,
      "paper_total_grad_release": 5,
      "pred_total_grad_release_with_scales": 5.125,
      "match_grad_release": true
    }
  },
  "tables_4_6_8": {
    "Table 4 Llama-3.1-8B AdamW": {
      "N_inferred_from_paper": 8026220134,
      "N_architecture": 8030261248,
      "N_rel_dev": 0.0005,
      "rows": {
        "Reference": {
          "paper": [
            29.9,
            59.8
          ],
          "pred": [
            29.915,
            59.83
          ],
          "max_rel_dev": 0.0005,
          "match": true
        },
        "FlashOptim": {
          "paper": [
            15.0,
            23.4
          ],
          "pred": [
            14.958,
            23.371
          ],
          "max_rel_dev": 0.0028,
          "match": true
        },
        "Weight Split": {
          "paper": [
            15.0,
            67.3
          ],
          "pred": [
            14.958,
            67.309
          ],
          "max_rel_dev": 0.0028,
          "match": true
        },
        "Opt. Quant.": {
          "paper": [
            29.9,
            15.9
          ],
          "pred": [
            29.915,
            15.892
          ],
          "max_rel_dev": 0.0005,
          "match": true
        }
      }
    },
    "Table 6 ResNet-50 SGD": {
      "N_inferred_from_paper": 26843545,
      "N_architecture": 25557032,
      "N_rel_dev": 0.0503,
      "rows": {
        "Reference": {
          "paper": [
            0.1,
            0.1
          ],
          "pred": [
            0.095,
            0.095
          ],
          "max_rel_dev": 0.0479,
          "match": true
        },
        "FlashOptim": {
          "paper": [
            0.05,
            0.05
          ],
          "pred": [
            0.048,
            0.049
          ],
          "max_rel_dev": 0.0479,
          "match": true
        },
        "Weight Split": {
          "paper": [
            0.05,
            0.12
          ],
          "pred": [
            0.048,
            0.119
          ],
          "max_rel_dev": 0.0479,
          "match": true
        },
        "Opt. Quant.": {
          "paper": [
            0.1,
            0.03
          ],
          "pred": [
            0.095,
            0.025
          ],
          "max_rel_dev": 0.157,
          "match": true
        }
      }
    },
    "Table 6 ResNet-50 AdamW": {
      "N_inferred_from_paper": 26843545,
      "N_architecture": 25557032,
      "N_rel_dev": 0.0503,
      "rows": {
        "Reference": {
          "paper": [
            0.1,
            0.19
          ],
          "pred": [
            0.095,
            0.19
          ],
          "max_rel_dev": 0.0479,
          "match": true
        },
        "FlashOptim": {
          "paper": [
            0.05,
            0.08
          ],
          "pred": [
            0.048,
            0.074
          ],
          "max_rel_dev": 0.0702,
          "match": false
        },
        "Weight Split": {
          "paper": [
            0.05,
            0.21
          ],
          "pred": [
            0.048,
            0.214
          ],
          "max_rel_dev": 0.0479,
          "match": true
        },
        "Opt. Quant.": {
          "paper": [
            0.1,
            0.05
          ],
          "pred": [
            0.095,
            0.051
          ],
          "max_rel_dev": 0.0479,
          "match": true
        }
      }
    },
    "Table 8 GPT-2 124M AdamW": {
      "N_inferred_from_paper": 123480309,
      "N_architecture": 124439808,
      "N_rel_dev": 0.0077,
      "rows": {
        "Reference": {
          "paper": [
            0.46,
            0.93
          ],
          "pred": [
            0.464,
            0.927
          ],
          "max_rel_dev": 0.0078,
          "match": true
        },
        "FlashOptim": {
          "paper": [
            0.23,
            0.36
          ],
          "pred": [
            0.232,
            0.362
          ],
          "max_rel_dev": 0.0078,
          "match": true
        },
        "Weight Split": {
          "paper": [
            0.23,
            1.04
          ],
          "pred": [
            0.232,
            1.043
          ],
          "max_rel_dev": 0.0078,
          "match": true
        },
        "Opt. Quant.": {
          "paper": [
            0.46,
            0.25
          ],
          "pred": [
            0.464,
            0.246
          ],
          "max_rel_dev": 0.0149,
          "match": true
        }
      }
    },
    "Table 8 GPT-2 124M Lion": {
      "N_inferred_from_paper": 123480309,
      "N_architecture": 124439808,
      "N_rel_dev": 0.0077,
      "rows": {
        "Reference": {
          "paper": [
            0.46,
            0.46
          ],
          "pred": [
            0.464,
            0.464
          ],
          "max_rel_dev": 0.0078,
          "match": true
        },
        "FlashOptim": {
          "paper": [
            0.23,
            0.24
          ],
          "pred": [
            0.232,
            0.239
          ],
          "max_rel_dev": 0.0078,
          "match": true
        },
        "Weight Split": {
          "paper": [
            0.23,
            0.58
          ],
          "pred": [
            0.232,
            0.579
          ],
          "max_rel_dev": 0.0078,
          "match": true
        },
        "Opt. Quant.": {
          "paper": [
            0.46,
            0.12
          ],
          "pred": [
            0.464,
            0.123
          ],
          "max_rel_dev": 0.0261,
          "match": true
        }
      }
    }
  },
  "checkpoint_claim": {
    "ref_bytes": 12,
    "flash_bytes": 5,
    "flash_with_scales": 5.125,
    "implied_7B_params": 7516192768,
    "pred_7B_flash_gib": 35.0,
    "paper_7B_flash_gib": 35,
    "reduction_pct": 58.3
  },
  "table1_all_match": true,
  "checkpoint_all_match": true,
  "tables_rows_matching": "19/20",
  "deviating_rows": {
    "Table 6 ResNet-50 AdamW / FlashOptim": {
      "paper": [
        0.05,
        0.08
      ],
      "pred": [
        0.048,
        0.074
      ],
      "max_rel_dev": 0.0702,
      "match": false
    }
  }
}